Abstract
"Classical" work in computer vision (e.g., Guzman, 1968) established that the grouping of surfaces to volumes in complex scenes—the key process in scene organization--could be achieved largely on the basis of constraints emanating from vertices where two or three contours co-terminate (Fig. 1). Surprisingly, with the focus on surface- or "appearance-based" models, there has been little research documenting the importance (or lack thereof) of vertices in shape-based object recognition, despite the extensive documentation that rapid object and scene perception can readily be achieved with line drawings. An L-vertex, the point at which two contours co-terminate, provides highly reliable evidence that a surface terminates at that vertex, thus providing the strongest constraint on the extraction of shape from images. In contrast, an X-vertex, where two contours cross without a change of direction at their crossing, provides no evidence for grouping and can thus be ignored in shape-based models of vision. 73 subjects named briefly presented line drawings of 56 objects in five conditions (Fig. 2): a) Intact (Original, O), b) OX, the Original with short line segments that crossed the contours to produce X-vertices, c) Contour Deleted (CD), the Original with gaps in each of the longer contours so that half of the original contour was deleted, d) CDX, the CD condition with short segments that crossed contours of the object thus producing X-vertices, e) CDL, the CDX condition with the segments shifted to the gap ends to produce L-vertices. Smooth continuation should allow grouping across the gaps, whether or not there were Xs present. Because the CDL condition bridges gaps with L vertices, each of which signals (inappropriately) the termination of a surface, it should interfere with the grouping of the object into a coherent whole, rendering identification difficult. This is what we observed (Fig. 3).
Meeting abstract presented at VSS 2016